Zeyu Xiong, Guangfan Sun
Accurate forecasting of urban electric vehicle (EV) charging demand is important for power system operation, sustainable transport electrification, and charging infrastructure planning. However, this task remains challenging because EV charging demand is shaped by temporal usage patterns as well as changing relationships among weather conditions, operational factors, and historical charging behavior. Many existing forecasting models treat these explanatory variables mainly as parallel inputs, while their mutual relationships are often predefined, simplified, or left implicit in the temporal learning process. To support AI-driven charging demand management, this study proposes an adaptive graph-based recurrent network (A-GRN) for city-level aggregated EV charging-load forecasting. In the proposed framework, key explanatory variables are represented as feature nodes, and their connections are learned through an adaptive adjacency matrix rather than a fixed spatial topology. The adaptive graph neural network (AGN) module captures feature-level interactions, while a dual-path gated recurrent unit module (DG-GRU) extracts temporal representations from the charging-load sequence. Experiments on a city-level EV charging dataset show that A-GRN outperforms several baseline models, including naive persistence forecasting, GRU, LSTM, BiGRU, TCN, and GCN. Compared with the BiGRU baseline, A-GRN reduces MAE, MSE, and RMSE by 31.36%, 34.65%, and 20.48%, respectively. In the original physical unit, the MAE is reduced from 187.43 kWh to 128.64 kWh, and the RMSE is reduced from 222.69 kWh to 177.08 kWh. The results indicate that feature-level graph learning can improve short-term EV charging-load forecasting, especially when the target is an aggregated urban load rather than the load of a single charging station. The proposed model provides a data-driven forecasting tool for sustainable urban charging demand management, low-carbon transport operation, and charging infrastructure planning.